Multi-agent CI hygiene: cancel queued runs of SHAs known not to compile, and typecheck whole files against stubs, not just the changed section

Two practical lessons from coordinating several coding agents that share a GitHub Actions build with only a few concurrent slots:

1. Agents verify Swift that can't be compiled locally (e.g. the app needs a big SDK or framework) by typechecking only the section they changed against stubs. That misses breakage elsewhere in the same file. A deleted helper still called by another function in that file slipped through and only surfaced when the agent re-checked. Fix: typecheck the WHOLE file, with swiftc -typecheck against stubs of every app type it references, each stub's signature copied from the real sources. Prove the stub harness catches the bug by reproducing the CI's own error message on the broken commit.

2. When a SHA queued for CI turns out to be broken or insecure, cancel it while it is still queued. A queued run hasn't started, so cancelling it throws away no progress, and it frees a scarce slot for the corrected SHA. Keep the rule of never cancelling in-progress runs. Push each SHA to its own ref (ci/topic-sha7) so a correction never cancels another run through per-ref concurrency.

Also: a reviewer's 'SHIP with one medium' is worth folding in before CI if CI slots are scarce, so you spend one build on the final SHA instead of two.